基于Faster R-CNN的破片群图像目标检测研究  

Target detection of warhead fragments group image based on Faster R-CNN

作　　者：雷江波 王泽民[1] 李静[1] Lei Jiangbo;Wang Zemin;Li Jing(School of Electronic Information Engineering,Xi’an University of Technology,Xi an 710021,China)

机构地区：[1]西安工业大学电子信息工程学院,西安710021

出　　处：《国外电子测量技术》2021年第1期70-74,共5页Foreign Electronic Measurement Technology

摘　　要：在研究战斗部战斗力与评价目标毁伤效能时,战斗部爆炸破片运动参数测试属于至为关键内容。破片的高速、小尺寸、多目标、发散性等特征和强火光烟尘环境使得破片群目标的检测和处理更有挑战性。经深入研究与探讨,提出了基于快速卷积神经网络(Faster R-CNN)的复杂背景下破片群检测法。破片图像通过Fast R-CNN的多层卷积和池化后得到特征图,由RPN根据特征图生成破片候选区域,再对破片候选区域进行池化,通过全连接层来预测破片边框位置,依靠区域生成网络(RPN)与Fast R-CNN共同训练所得网络完成破片群目标检测。经过对复杂背景下不同破片数量的真实高速图像进行实验,证明该方法可从中自动提取破片图像深层特征,破片群目标检测的准确率达到96%,平均识别率为93.4%。The motion parameters testing of warhead fragments have significant influence for ammunition lethality and target vulnerability assessment.The characteristics of fragment group,such as high speed,small size,multi-target,divergence and strong fire smoke environment,give a challenge to fragment image processing.The novel method based on Faster R-CNN is presented to detect the fragments images.The images are convolved and pooled by Fast R-CNN to obtain feature maps.The fragments candidate regions are extracted by RPN according to the feature maps,and pooled by the pooling layer to extract deep and abstract features of fragmentation.The full connection layer is used to predict the fragmentation border position.The jointly trained network RPN and Fast R-CNN detect fragments target.The experimental results show that the method can extract fragments features for the warhead fragments images.The detection accuracy and average recognition rate are 96%and 93.4%respectively.

关 键 词：战斗部破片检测 高速视觉成像 多目标检测 Faster R-CNN 区域生成网络 

分 类 号：TN98[电子电信—信息与通信工程]